Pump room intelligent scheduling control working method and system based on machine learning, and medium

Abstract

The invention provides a pump room intelligent scheduling control working method and system based on machine learning and a medium, and the method comprises the following steps: S1, inputting a target flow, and predicting the pressure and flow working conditions of a corresponding header pipe; s2, deducing and screening a single-pump flow distribution strategy and a single-pump water outlet pressure strategy which meet the main pipe working condition; s3, predicting the water pump starting frequency required by the flow distribution of each single pump and the water outlet pressure strategy of each single pump; and S4, predicting the energy consumption condition of the starting strategy of each water pump, and screening out the scheme with the lowest energy consumption as a final control scheme. According to the method, the main pipe flow pressure working condition prediction model based on the neural network is introduced, so that the pressure and the flow of each water outlet main pipe under the target flow can be accurately predicted. The method avoids the dependence of a traditional control method on experience, can dynamically adapt to different pipeline characteristics and water plant requirements in real time, and ensures the accuracy of system working condition prediction.

Description

Technical Field

[0001] The present invention relates to the technical field of water dispatching, and in particular to a method, system and medium for intelligent dispatching control of a pump room based on machine learning. Background Art

[0002] In the traditional raw water pump room control system, the operation of the pump usually relies on manual experience or a fixed mode based on a simple control algorithm. Common methods include fixed speed operation and controlling the water volume by adjusting the valve opening. Although this method is simple, it often wastes energy and cannot efficiently cope with the limitations of dynamic water volume demand. With the increasing complexity of the pump room system operation, the expansion of the pipeline network, and the real-time fluctuations in flow and pressure, it is difficult for traditional control methods to take into account the economic operation of the pump and the dynamic needs of the pipeline network. In order to improve efficiency, some pump rooms have begun to introduce pump control methods based on variable frequency technology, allowing pumps to flexibly adjust frequency under different load conditions to achieve energy saving. However, most of the current variable frequency control methods rely on simple PID control or rule control of a single pump, lacking comprehensive consideration of the overall system complexity, operating stability, and energy consumption optimization, and cannot achieve the overall optimal pump control strategy. In recent years, intelligent control methods based on machine learning and optimization algorithms have gradually been applied to the field of pump room control. These methods analyze historical operating data, build mathematical models to predict the system's flow, pressure and other operating conditions, and use optimization algorithms to screen out the optimal control strategy. However, these methods still have shortcomings in dealing with complex pipeline networks and multi-pump parallel systems in terms of operating stability judgment and energy consumption optimization. They usually rely on fixed models or simple linear relationships and cannot achieve precise control of the dynamic conditions of the pump group and minimum energy consumption operation. Summary of the invention

[0003] The present invention aims to at least solve the technical problems existing in the prior art, and in particular innovatively proposes a pump room intelligent dispatching control working method, system and medium based on machine learning.

[0004] In order to achieve the above-mentioned object of the present invention, the present invention provides a pump room intelligent dispatching control working method based on machine learning, comprising the following steps:

[0005] S1, input the target flow rate and predict the corresponding main pipe pressure and flow conditions;

[0006] S2, deduce and select the single pump flow distribution and single pump outlet pressure strategy that meets the main pipe working conditions;

[0007] S3, predicting the pump activation frequency required for each single pump flow distribution and single pump water outlet pressure strategy;

[0008] S4, predict the energy consumption of each pump activation strategy and select the plan with the lowest energy consumption as the final control plan.

[0009] In a preferred embodiment of the present invention, step S1 includes:

[0010] S11, obtaining current pipeline characteristic related data and target flow data;

[0011] S12, obtaining current pipeline characteristic related data and target flow data and inputting them into the main pipe flow pressure condition prediction model to obtain the pressure and flow distribution of each outlet main pipe under the target flow.

[0012] In a preferred embodiment of the present invention, in step S11, the current pipeline characteristic related data includes the flow rate Q of each outlet main pipe of the current raw water pump room. c 、Pressure P c , the opening degree of each valve in the pipeline c , the inflow flow Q of the downstream water plant out One or any combination.

[0013] In a preferred embodiment of the present invention, step S12 includes:

[0014] The main pipe flow and pressure condition prediction model uses a neural network structure with 2 hidden layers, 32 neurons in each hidden layer, and ReLU as the activation function; it is expressed as:

[0015]

[0016] Among them, f1(·) is the neural network model;

[0017] θ1 is the parameter of the model;

[0018] W 1,1 , W 1,2 , W 1,3 is the weight matrix;

[0019] b 1,1 、b 1,2 、b 1,3 is the bias vector;

[0020] When training the model, the historical running data is used as the training set, all input and output data are normalized, and the mean square error is used as the loss function:

[0021]

[0022] Where N is the number of training samples;

[0023] Q a,i and P a,i is the total pipe flow and pressure of the i-th sample;

[0024] and is the predicted value of total pipe flow and pressure of the i-th sample output by the model;

[0025] in and are the main outlet water flow and pressure predicted by the model respectively.

[0026] In a preferred embodiment of the present invention, step S2 includes:

[0027] A neural network structure with 2 hidden layers is used, each hidden layer has 32 neurons, the activation function is ReLU, and the activation function of the output layer is sigmoid, which is expressed as

[0028] S = f2(X2;θ2) = Sigmoid(W 2,3 (ReLu(W 2,2 (ReLu(W 2,1 X1+b 2,1 ))+b 2,2 ))+

[0029] b 2,3 ),

[0030] Among them, f2(·) is the neural network model;

[0031] θ2 is the parameter of the model;

[0032] W 2,1 , W 2,2 , W 2,3 is the weight matrix;

[0033] b 2,1 、b 2,2 、b 2,3 is the bias vector;

[0034] When constructing the data set, for the operating condition data during historical stable operation, the label is set to positive example y=1; the label is set to negative example y=0; and the binary cross entropy loss function is used to complete the training of the operating condition stability judgment model:

[0035]

[0036] Where N is the number of training samples;

[0037] y i is the label of the i-th sample;

[0038] S i It is the stability prediction value output by the model.

[0039] In a preferred embodiment of the present invention, step S3 includes:

[0040] A neural network structure with 2 hidden layers is used, with 32 neurons in each hidden layer, ReLU as the activation function, and mean square error as the loss function; it is expressed as

[0041]

[0042] Among them, f 3,i (·) is the neural network model of the ith single pump;

[0043] θ 3,i are the parameters of the model;

[0044] W 3,1 , W 3,2 , W 3,3 is the weight matrix;

[0045] b 3,1 、b 3,2 、b 3,3 is the bias vector;

[0046] The historical real operation data is used to form the input and output of the training data set, and the loss function uses the mean square error, the formula is:

[0047]

[0048] Among them, F i,j is the true historical frequency of the jth sample of the i-th water pump;

[0049] The frequency predicted by the model.

[0050] In a preferred embodiment of the present invention, step S4 includes:

[0051] A neural network structure with 2 hidden layers is used, with 32 neurons in each hidden layer, ReLU as the activation function, and mean square error as the loss function, expressed as

[0052]

[0053] Among them, f 4,i (·) is the neural network model of the ith single pump;

[0054] θ 4,i are the parameters of the model;

[0055] W 4,1 , W 4,2 , W 4,3 is the weight matrix;

[0056] b 4,1 、b 4,2 、b 4,3is the bias vector;

[0057] The historical real operation data is used to form the input and output of the training data set, and the loss function uses the mean square error, the formula is:

[0058]

[0059] Among them, E i,j is the real historical power of the jth sample of the i-th water pump;

[0060] is the power predicted by the model.

[0061] The present invention also discloses a computer system, comprising:

[0062] processor;

[0063] a memory for storing processor-executable instructions;

[0064] Among them, the processor is configured to implement the pump room intelligent scheduling and control working method based on machine learning when executing the executable instructions.

[0065] The present invention also discloses a computer-readable storage medium, comprising:

[0066] a memory having a computer program stored thereon;

[0067] A processor is used to execute the program in the memory to implement the pump room intelligent scheduling and control working method based on machine learning.

[0068] In summary, by adopting the above technical solution, the present invention can achieve:

[0069] 1) Accurate working condition prediction: By introducing a neural network-based main pipe flow pressure working condition prediction model, the pressure and flow of each outlet main pipe at the target flow rate can be accurately predicted. This method avoids the reliance of traditional control methods on experience, can dynamically adapt to different pipeline characteristics and water plant needs in real time, and ensure the accuracy of system working condition prediction.

[0070] 2) Intelligent flow distribution and working condition stability judgment: Through the working condition stability judgment model, combined with the search of single pump flow and pressure combinations, the single pump flow distribution strategy that meets the main pipe working condition can be intelligently screened. Compared with the traditional water pump control method, the present invention uses a machine learning model to realize the automatic adjustment of the pump system, ensuring the stable operation of the system under complex working conditions and greatly reducing manual intervention and errors.

[0071] 3) Energy consumption optimization: The power consumption under different single pump strategies is evaluated through the operating power prediction model. The system can automatically select the pump operation plan with the lowest energy consumption, ensuring that the pump room system minimizes energy consumption while meeting the target flow demand, thereby improving overall operating efficiency.

[0072] 4) Efficient management and decision support: Through continuous learning of historical data, the system can gradually optimize the operation strategy of the water pump to adapt to different changes in pipe network conditions and water plant needs. At the same time, the automated control strategy can respond quickly under different operating conditions, reduce the lag of manual decision-making, and improve the intelligence level of pump room management.

[0073] Additional aspects and advantages of the present invention will be given in part in the following description and in part will be obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0074] The above and / or additional aspects and advantages of the present invention will become apparent and easily understood from the description of the embodiments in conjunction with the following drawings, in which:

[0075] Figure 1 It is a schematic diagram of the process of the present invention. DETAILED DESCRIPTION

[0076] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and cannot be understood as limiting the present invention.

[0077] Combination Figure 1 , specifically including the following steps:

[0078] Step 1: Input the target flow rate and predict the corresponding main pipe pressure and flow conditions.

[0079] A raw water pump house usually delivers water of a specified flow rate to multiple water plants downstream. When the valve state in the pipeline system from the raw water pump house to the downstream water plant is selected, the flow rate and pressure of the outlet main pipe of the raw water pump house are in a one-to-one relationship, and the main pipe pressure increases with the increase of the flow rate. The present invention determines the pressure and flow distribution of each outlet main pipe under the target flow rate through a main pipe flow pressure working condition prediction model. Specifically, the input of the model mainly includes two types of information:

[0080] 1) Current pipeline characteristics related data: including the flow rate Q of each outlet pipe in the current raw water pump room c 、Pressure P c , the opening of each valve in the pipeline is O c, the inflow flow rate Q of the downstream water plant out These data are used to determine the characteristics of the pipeline.

[0081] 2) Target flow data: the total water flow rate expected by the raw water pump room (Q t ).

[0082] The input of the model can be expressed as X1 = [Q c ,P c ,O c ,Q out ,Q t The output of the model is the target flow Q t Flow rate Q of each outlet main pipe of the lower raw water pump room a 、Pressure P a (When there is only one main water outlet Q t =Q a ).

[0083] The main pipe flow and pressure condition prediction model uses a 2-hidden-layer neural network structure, with 32 neurons in each hidden layer and ReLU as the activation function. The specific neural network calculation can be expressed as

[0084]

[0085] Among them, f1(·) is the neural network model, θ1 is the parameter of the model, and W 1,1 , W 1,2 , W 1,3 is the weight matrix, b 1,1 、b 1,2 、b 1,3 is the bias vector.

[0086] When training the model, the historical running data is used as the training set, all input and output data are normalized, and the mean square error (MSE) is used as the loss function:

[0087]

[0088] Among them, N is the number of training samples, Q a,i and P a,i is the total pipe flow and pressure of the i-th sample, and It is the predicted value of total pipe flow and pressure of the i-th sample output by the model.

[0089] in and They are the main pipe outflow and pressure predicted by the model. The adaptive optimization algorithm Adam is used for training and updating.

[0090] When implementing step 1, the pipeline characteristic related data of the current pump room operating environment and the target flow data are input into the main pipe flow pressure condition prediction model to obtain the pressure and flow distribution of each outlet main pipe under the target flow.

[0091] Step 2: deduce and select the single pump flow distribution and single pump water outlet pressure strategies that meet the main pipe working conditions.

[0092] The raw water pump room is usually supplied with water by multiple water pumps in parallel, that is, the water pipes of multiple single pumps are gathered into a main pipe, and then the water is supplied to different directions out of the factory. In the process of clean water reaching the outlet main pipe from the pump, the head friction loss caused by the valves in the pipeline usually being kept fully open is minimized. Therefore, the present invention only considers the pump strategy when the valves in the source water pump room are fully open, and makes full use of the frequency conversion function of the water pump to achieve the water supply demand under all target flow rates. When the pump group is running, after the flow and pressure of each single pump are stable, the flow condition of the main pipe will also be in a stable state. Therefore, the present invention constructs a working condition stability judgment model to deduce whether a group of single pump flow pressures and the main pipe flow pressure will be in a stable operating state. Specifically, the input and output of the working condition stability judgment model are:

[0093] 1) Input: The water flow rate Q of each single pump p 、Pressure P p , the flow rate Q of each pump room outlet pipe a 、Pressure P a .

[0094] 2) Output: Is it in stable operating state S?

[0095] Specifically, suppose there are N pump rooms. p Single pump, total water outlet pipe is N a Root, then the input of the model is Where Q p,i is the water flow rate of the ith single pump, P p,i is the outlet pressure of the i-th single pump, Q a,j is the outflow rate of the jth main pipe, P a,j is the outlet pressure of the jth main pipe. The output of the model is the stability judgment value S, which is a floating point number in the range of [0,1]. The closer it is to 1, the more stable the state is, and the closer it is to 0, the more unstable it is.

[0096] The model uses a neural network structure with 2 hidden layers, 32 neurons in each hidden layer, ReLU activation function, and sigmoid activation function in the output layer. The specific neural network calculation can be expressed as

[0097] S = f2(X2;θ2) = Sigmoid(W 2,3 (ReLu(W 2,2(ReLu(W 2,1 X1+b 2,1 ))+b 2,2 ))+b 2,3 )

[0098] Where f2(·) is the neural network model, θ2 is the parameter of the model, and W 2,1 , W 2,2 , W 2,3 is the weight matrix, b 2,1 、b 2,2 、b 2,3 is the bias vector.

[0099] When constructing the data set, the label of the working condition data during historical stable operation is set to positive example y=1; in addition, the label of the data after adding random noise is set to negative example y=0. The binary cross entropy loss function is used to complete the training of the working condition stability judgment model:

[0100]

[0101] Among them, N is the number of training samples, y i is the label of the i-th sample, S i The stability prediction value of the model output. The adaptive optimization algorithm Adam is used for training and updating.

[0102] When implementing step 2, various possible single pump flow and pressure combination schemes and the target main pipe flow and pressure obtained in step 1 are input into the working condition stability judgment model, and the scheme that satisfies the main pipe working condition is set as the model output that satisfies S≥τ (τ is usually a decimal between 0.85 and 1). This can screen out the single pump flow distribution and single pump outlet pressure strategies that meet the target flow. Since both pressure and flow are continuous data with a huge search space, it can be achieved by setting the pressure and flow intervals (for example, setting the pressure interval to 0.001Mpa and the flow rate to 10m 3 / h), converting it into discrete data so that the model can explore all feasible solutions. When implementing this step, various constraints can be added to reduce the possibility space, such as the rotation time limit of the pump, some pumps must be turned off or must be enabled, etc.

[0103] After step 2, all single pump pressure and flow conditions that meet the target main pipe conditions can be obtained.

[0104] Step 3: predict the pump activation frequency required for each single pump flow distribution and single pump water outlet pressure strategy.

[0105] For each single pump, when the outlet flow, pressure and suction port liquid level are clear, that is, the flow and head are determined, only a specified unique frequency can achieve the working condition. Therefore, the present invention constructs a working condition activation frequency prediction model for each single pump, and inputs the suction liquid level height (h i )、Outlet pressure (P p,i ) and water flow rate (Q p,i ), output the operating frequency of the pump (F i ). The input vector can be represented as X p,i =[h i ,P p,i ,Q p,i ]

[0106] The model uses a neural network structure with 2 hidden layers, 32 neurons in each hidden layer, ReLU as the activation function, and mean square error (MSE) as the loss function. The calculation of the neural network model can be expressed as

[0107]

[0108] Among them, f 3,i (·) is the neural network model of the ith single pump, θ 3,i is the parameter of the model, W 3,1 , W 3,2 , W 3,3 is the weight matrix, b 3,1 、b 3,2 、b 3,3 is the bias vector.

[0109] The historical real operation data is used to form the input and output of the training data set, and the loss function uses the mean square error (MSE), the formula is:

[0110]

[0111] Among them, F i,j is the real historical frequency of the jth sample of the i-th water pump, The frequency predicted by the model. The adaptive optimization algorithm Adam is used for training updates.

[0112] When implementing step 3, input the single pump pressure and flow obtained in step 2, as well as the real-time water absorption level collected in the pump room, to obtain the operating frequency of the single pump.

[0113] Step 4: predict the energy consumption of each pump activation strategy and select the plan with the lowest energy consumption as the final control plan.

[0114] For each single pump, when the flow pressure at the outlet and the suction level and the operating frequency are clear, that is, the flow head and the frequency are determined, the power required can be estimated. The present invention constructs a working condition power prediction model for each single pump, and inputs the outlet pressure (P p,i ), flow rate (Q p,i ), water absorption level height (h i )、Operation frequency (F i ), output power of the pump (E i ). The input vector can be represented as X E,i =[h i ,P p,i ,Q p,i ,F i ]

[0115] The model uses a neural network structure with 2 hidden layers, 32 neurons in each hidden layer, ReLU as the activation function, and mean square error (MSE) as the loss function. The calculation of the neural network model can be expressed as

[0116]

[0117] Among them, f 4,i (·) is the neural network model of the ith single pump, θ 4,i is the parameter of the model, W 4,1 , W 4,2 , W 4,3 is the weight matrix, b 4,1 、b 4,2 、b 4,3 is the bias vector. Adaptive optimization algorithm Adam is used for training updates.

[0118] The historical real operation data is used to form the input and output of the training data set, and the loss function uses the mean square error (MSE), the formula is:

[0119]

[0120] Among them, E i,j is the real historical power of the jth sample of the i-th pump, is the power predicted by the model.

[0121] When implementing step 4, the single pump operating condition that can achieve the target flow obtained in step 3 is input into the operating condition power prediction model to obtain the power of the operating strategies of all pump groups. The single pump operating power of each strategy is summed to obtain its expected total pump group power. Where N P is the number of single pumps. The strategy with the minimum expected total power is selected and its pump frequency operation combination is used as the final control solution.

[0122] Although the embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the claims and their equivalents.

Claims

1. A pump room intelligent dispatching control working method based on machine learning, characterized in that: The following steps are involved: S1, input the target flow rate and predict the corresponding main pipe pressure and flow conditions; S2, deduce and select the single pump flow distribution and single pump outlet pressure strategy that meets the main pipe working conditions; S3, predicting the pump activation frequency required for each single pump flow distribution and single pump water outlet pressure strategy; S4, predict the energy consumption of each pump activation strategy and select the plan with the lowest energy consumption as the final control plan.

2. The pump room intelligent dispatching control working method based on machine learning according to claim 1 is characterized in that: Step S1 includes: S11, obtaining current pipeline characteristic related data and target flow data; S12, obtaining current pipeline characteristic related data and target flow data and inputting them into the main pipe flow pressure condition prediction model to obtain the pressure and flow distribution of each outlet main pipe under the target flow.

3. The pump room intelligent dispatching control working method based on machine learning according to claim 1 is characterized in that: In step S11, the current pipeline characteristic related data includes the flow rate Q of each outlet main pipe of the current raw water pump room c 、Pressure P c , the opening of each valve in the pipeline c , the inflow flow Q of the downstream water plant out One or any combination.

4. The pump room intelligent dispatching control working method based on machine learning according to claim 1 is characterized in that: Step S12 includes: The main pipe flow and pressure condition prediction model uses a neural network structure with 2 hidden layers, 32 neurons in each hidden layer, and ReLU as the activation function; it is expressed as: Among them, f1(·) is the neural network model; θ1 is the parameter of the model; W 1，1 , W 1，2 , W 1，3 is the weight matrix; b 1，1 、b 1，2 、b 1，3 is the bias vector; When training the model, the historical running data is used as the training set, all input and output data are normalized, and the mean square error is used as the loss function: Where N is the number of training samples; Q a，i and P a，i is the total pipe flow and pressure of the i-th sample; and is the predicted value of total pipe flow and pressure of the i-th sample output by the model; Its look and are the main outlet water flow and pressure predicted by the model respectively.

5. The pump room intelligent dispatching control working method based on machine learning according to claim 1 is characterized in that: Step S2 includes: A neural network structure with 2 hidden layers is used, each hidden layer has 32 neurons, the activation function is ReLU, and the activation function of the output layer is sigmoid, which is expressed as S=f2(X2;θ2)=Sigmoid(W 2，3 (ReLu(W 2，2 (ReLu(W 2，1 X1+b 2，1 ))+b 2，2 ))+b 2,3 ), Among them, f2(·) is the neural network model; θ2 is the parameter of the model; W 2,1 , W 2,2 , W 2，3 is the weight matrix; b 2,1 、b 2,2 、b 2,3 is the bias vector; When constructing the data set, for the operating condition data during historical stable operation, the label is set to positive example y=1; the label is set to negative example y=0; and the binary cross entropy loss function is used to complete the training of the operating condition stability judgment model: Where N is the number of training samples; y i is the label of the i-th sample; S i It is the stability prediction value output by the model.

6. The pump room intelligent dispatching control working method based on machine learning according to claim 1 is characterized in that: Step S3 includes: A neural network structure with 2 hidden layers is used, with 32 neurons in each hidden layer, ReLU as the activation function, and mean square error as the loss function; it is expressed as Among them, f 3，i (·) is the neural network model of the ith single pump; θ 3，i are the parameters of the model; W 3，1 , W 3，2 , W 3，3 is the weight matrix; b 3，1 、b 3，2 、b 3，3 is the bias vector; The historical real operation data is used to form the input and output of the training data set, and the loss function uses the mean square error, the formula is: Among them, F i，j is the true historical frequency of the jth sample of the i-th water pump; The frequency predicted by the model.

7. The pump room intelligent dispatching control working method based on machine learning according to claim 1 is characterized in that: Step S4 includes: A neural network structure with 2 hidden layers is used, with 32 neurons in each hidden layer, ReLU as the activation function, and mean square error as the loss function, expressed as Among them, f 4，i (·) is the neural network model of the ith single pump; θ 4，i are the parameters of the model; W 4，1 , W 4，2 , W 4，3 is the weight matrix; b 4，1 、b 4，2 、b 4，3 is the bias vector; The historical real operation data is used to form the input and output of the training data set, and the loss function uses the mean square error, the formula is: Among them, E i，j is the real historical power of the jth sample of the i-th water pump; is the power predicted by the model.

8. A computer system, characterized in that: include: processor; a memory for storing processor-executable instructions; Wherein, the processor is configured to implement the pump room intelligent dispatching and control working method based on machine learning as described in one of claims 1 to 7 when executing the executable instructions.

9. A computer-readable storage medium, characterized in that: include: a memory having a computer program stored thereon; A processor is used to execute the program in the memory to implement the pump room intelligent scheduling and control working method based on machine learning as described in any one of claims 1 to 7.

Images